Mining and Injecting Legal Prior Knowledge to Improve the Generalization Ability of Neural Networks in Chinese Judgments

Yaying Chen, Ji Ma, Nanfei Gu, Xintong Wang, Minghua He

Abstract: Data mining often faces the problem of too many missing values in real datasets and of the inability to quantify the value of mined knowledge, which constrains the generalization ability of neural networks. In publicly available legal datasets, established data filling methods ignores the invisible constraints among non-random missing data. To address this issue, optimal Bayesian network and posterior probability are adopted to predict missing values, which can estimate the true distribution of missing variables through mining the implicit constraints in existing available variable data. At the same time, the average treatment effect on the treated (ATT) is adopted to mine the adjudication paths in legal datasets which are injected into Bi-LSTM as prior knowledge. We have demonstrated through extensive experiments that the injection of prior knowledge is effective in improving the generalization ability of neural networks. In addition, through the Legal Judgment Prediction (LJP) task, the value of the mined prior knowledge can be evaluated, the most critical adjudication path can be identified which contributes to the promotion of judicial uniformity.